AA-3D Benchmarks AI Models That Build Interactive 3D Experiences

AA-3D Benchmarks AI Models That Build Interactive 3D Experiences

AA-3D evaluates how effectively AI models create interactive 3D scenes that people want to explore. Each model operates as an agent inside an offline sandbox, converting a written brief into a three.js web page.

A recruited panel compares pairs of scenes created from the same brief without knowing which model produced each scene. The results are used to calculate Elo ratings. In the AA-3D v1.0 results recorded on October 12, 2026, Claude Sonnet 5.5 ranks first with an Elo score of 1194. Its score is statistically tied with three other Anthropic and OpenAI models.

AA-3D v1.0 leaderboard

Elo scores are based on blind pairwise votes, with 95% confidence intervals. Higher scores indicate better performance.

Claude Sonnet 5.5 has the highest score at 1194, although the top four models are statistically tied because their confidence intervals overlap. All four models come from Anthropic and OpenAI.

Meta’s Muse Spark 1.3 ranks seventh with 1035 Elo. The highest-ranked open weights models are Kimi K3, at 982, and MiMo-V2.6-Pro, at 977. They rank 11th and 12th, respectively, more than 200 points behind the leader.

These are the launch results for AA-3D v1.0. The leaderboard can change as additional models are evaluated.

How AA-3D works

Each model receives a written brief describing an interactive scene and a shared delivery contract. It then works as an agent in an offline sandbox, where it can write code, render the scene, inspect screenshots of its work, and revise the result.

The sandbox includes:

• Blender
• Rapier physics
• Python
• A headless browser
• A library of CC0 assets

Each run ends after three hours or 250 turns. The output is an index.html page containing an interactive scene that runs in a browser.

Scene-building timelines and costs

GPT-6 Astra produces its first complete page after six minutes and finishes after 68 minutes. Claude Sonnet 5.5 produces no visible output until minute 42 and continues working until two hours and 58 minutes.

Claude Sonnet 5.5 generates 6.2 times as many output tokens as GPT-6 Astra, while costing 1.3 times as much. Its run costs $14.93, compared with $11.62 for GPT-6 Astra. The list prices for Claude Sonnet 5.5 are one-fifth of those for GPT-6 Astra.

A complete AA-3D scene costs between $0.20 and $32.08, depending on the model. GPT-6 Luna is the least expensive at $0.20 per scene, while Claude Fable 5.1 is the most expensive. Input accounts for 69% of total spending because agents resend their full context on every turn.

Among the five highest-ranked models, GPT-6 Sol has the lowest cost at $3.46 per scene. Claude Sonnet 5.5 costs 4.4 times as much, and Claude Fable 5.1 costs 9.3 times as much.

GPT-6 Luna builds all 55 scenes and still costs only $0.20 per scene while ranking eighth. Kimi K3 costs 68 times as much per scene for a similar Elo score.

What the benchmark shows about 3D generation

Higher-ranked models inspect their work earlier

The five highest-ranked models first inspect an image of their own scene after a median of 10 to 13 turns. In at least 89% of their scenes, they do so by turn 20. Models ranked 14th through 20th wait until turns 15 to 40.

Among the 19 models that can view images, early visual inspection has a strong relationship with Elo, with a Spearman correlation of 0.82. The top five models also begin browser testing sooner, at a median of 16 turns compared with 25 turns for the others.

How early a model looks appears more important than how often it looks. MiniMax-M3 inspects its own scene 54 times per scene, more than any other model, but views its work by turn 20 in only 7% of scenes. It ranks 19th.

GLM-5.3 and Nemotron 3 Ultra are text-only models and built their scenes without an image-viewing tool.

Detailed scenes receive more votes

The most detailed fifth of scenes within each task produces screenshots that are twice as large, 168 KB compared with 84 KB, and has one-third as much empty space, 22% compared with 60%.

These scenes win 63% of votes, compared with 31% for the least detailed fifth. Model Elo alone predicts 61% and 33%, respectively, indicating that much of the difference is associated with stronger models. The data does not establish whether detail itself contributes to a model’s overall strength.

Within a single model, the effect is smaller. A scene that is one standard deviation more detailed than that model’s other scenes wins approximately 1.5 additional percentage points of its votes, with estimates ranging from 1.3 to 1.8 depending on the detail measure.

Of 86 scene measures, 11 remain significant after correction. These include detail measures, render resolution, antialiasing, GPU time per frame, and the number of page elements. None changes vote share by more than four points.

OpenAI’s GPT-6 models outperform their general-intelligence predictions

General intelligence is strongly associated with AA-3D performance. Each point on the Artificial Analysis Intelligence Index corresponds to approximately 18 Elo points, with a Pearson correlation of 0.92. This is a closer relationship than the correlations for Terminal-Bench 4.0, at 0.81, or SciCode, at 0.76.

GPT-6 Luna and GPT-6 Sol exceed this trend by 141 and 93 Elo points, respectively. GPT-6 Astra exceeds it by 46 points. GLM-5.3 falls furthest below the trend, by 197 points.

GLM-5.3 cannot view images, which may account for part of the difference. Z AI’s GLM-5.3-Flash can view images and scores 149 points higher despite having a lower Intelligence Index. Nemotron 3 Ultra is also text-only and sits close to the predicted trend.

Asset-library use does not improve vote results

Three out of four scenes use an item from the CC0 asset library, whose use is encouraged by the shared instructions. Models that use the library more frequently tend to rank higher, with a Spearman correlation of 0.53.

However, after accounting for the opponent, a model’s library-based scenes do not win votes more often than its other scenes. GPT-6 Luna uses the library in only one-third of its scenes and ranks eighth overall.

Tokens, turns, and tool use

Models that generate more text do not necessarily score higher. Output-token count and Elo have no clear relationship, with a Spearman correlation of 0.33 that is not statistically significant.

Claude Sonnet 5.5 generates the most output, averaging 569,000 output tokens per scene, approximately two-thirds of which are reasoning tokens. GPT-6 Sol generates 109,000 output tokens and ranks fifth. Input tokens account for 97% to 99% of all tokens, and most models read nearly all input from the cache.

Turn count also has no clear relationship with Elo. The Spearman correlation is -0.27 and is not statistically significant. GLM-5.3 reaches the 250-turn limit in 64% of its scenes. GPT-6.1 Sol and GPT-6 Astra require median totals of only 56 and 60 turns, but each turn lasts approximately two minutes, most of that time being spent by the model.

Agents spend 82% of their tool time testing the page in a browser. Browser tests, which either capture a screenshot or check for errors, make up only one-third of all tool calls. Because each test takes 33 to 38 seconds, browser testing accounts for about 25 of the 30.7 minutes of tool time per scene.

Code writing accounts for one tool call in six and only 6% of tool time. The five highest-ranked models read files and documentation about half as often as the other models, 9.4 times per scene compared with 18.7 times. Across all 21 models, reading less is associated with a higher Elo, with a Pearson correlation of -0.74.

Interactive controls and performance

The highest-ranked models generally place more interactive elements on the page. The top five have a median of 15 to 20 buttons and other controls per scene, compared with six to eight for the bottom three models. Control count has a strong relationship with Elo, with a Pearson correlation of 0.90.

Rendering speed does not distinguish the models. Every model has a median frame time below 1.3 milliseconds of GPU time on the test machine, well below the 16.7 milliseconds available for a 60-frame-per-second experience.

Performance across domains

The same models tend to perform well across nearly every domain. A top-five model has the highest win rate in six of the seven domains.

In the living-bodies domain, GPT-6.1 Sol, which ranks sixth overall, and Claude Sonnet 5.5 are effectively tied at the top, separated by less than one percentage point.

The top five also perform consistently across domains. In every domain, each remains within 12 points of the win rate predicted by its overall Elo against the opponents it encountered there. Only six of the other models’ 112 domain cells differ by more than that amount, and three of those cells are marked as having limited data.

Each cell is based on two to ten scenes, with a median of seven. Cells containing fewer than five scenes are excluded, and differences of approximately 12 points or less are considered within the range of noise.

Inside an agent run

Across six tasks, GPT-6 Sol is the fastest of the five highest-ranked models. Its median run takes 60 minutes, compared with 115 to 151 minutes for the other four.

The top five do not spend more time testing than the other models. As a group, they spend 25% of their time on browser tests, approximately the same share as the full set of models.

Every top-five run includes at least one view of the model’s own scene. GLM-5.3 and Nemotron 3 Ultra are text-only, so none of their 12 runs includes an image inspection.

AA-3D resources

The AA-3D leaderboard contains the full results and can be updated as new models are released. The methodology documents the generation settings, scene checks, and Elo calculations. AA-3D runs on Stirrup, an open-source agent harness.